Methodology to Minimize Risk in SME Lending for the US and India Markets

Abstract

Fintech start-ups offer digital SME Lending Platforms for smooth customer journeys, that spans from campaigning to loan delivery to small enterprises. These startups are rising in numbers across the globe specifically the United States and India Markets due to the increased demand of the credit by the small and medium enterprises. The widely used phrase "lending" refers to the process of lender giving money to the borrower.
Digitalization uses state through processing to bring cost-effectiveness, enhanced customer experience, speed of disbursements, better risk assessment, and adequate decision-making.
SME, or small and medium-sized enterprises, are companies that need capital to operate in the market. To grow their business, they require capital. Since the majority of them are new companies, they fall into the subprime or thin-file consumer category, which means that credit agencies have either given them a low credit score or no credit score at all. These customers' loan requests are denied by financial institutions because they are unable to meet their capital criteria. By leveraging alternative data such as payments, rent, mobile phone, and social media information, the fintech start-up industries target these customers
and offer loans at higher interest rates in an effort to reduce the risk. They have also developed their own digital systems for risk assessment of these customers.
This paper provides the improvisation pertaining to the Bank Statement Analyser built in the study (Mehta, 2020) to mitigate the risk associated with the SME lending in the United States and India. At the same time this method uses the latest AI technology via Large Language Models that has a potential to boost the credit decisioning process of a financial institution.
The statements that banks periodically release are known as bank statements. Transactionlevel information for their clients is included in these statements. This study examined 1,00,000 transactions from 3,300 small and medium-sized businesses in the US and 20,000 transactions from 475 small and medium enterprises in the India market using anonymized parsed PDF bank statement data. Using text mining techniques like count vectorizer and ngrams, as detailed in the study (Mehta, 2020), this transaction-level data was first used to generate transaction categories, or keywords. Later, these keywords were used to classify
the transactions. Compared to earlier research (Mehta, 2020), a more refined set of keywords is used in this study. In this work, rule-based, machine learning, and Large Language Models methods for multi-class classification were employed as classification
techniques. Various prompt designs and strategies are used for large language models classification task.
In the end, a number of important measures known as cash flow variables were generated for loan credit decisioning, together with changes in revenue, operational profit, operating profit margin, and debt services.
KEYWORDS: LLM in Banking, Bank Statements Analysis, Machine Learning in Lending, Digital SME Lending, Digital Lending, Risk in SME Lending, AI to reduce risk in SME Lending, Use of large language models for credit decisioning